Humanizing GPT-3: Introducing OCEAN
I’m a researcher, not a programmer. Don’t walk away from this paper disappointed by the lack of a detailed technical explanation or concrete examples of the ideas within. I have never produced text written by GPT-3. What attracts me to the technology is the potential to combine personality and technology along the lines of the English language. Fundamentally, GPT-3 is a genius AI designed to produce English text that passes the Turing test. People are excited by its ability to predict and mimic, but in it I see the potential for so much more.
Let me introduce you to the Big 5, or OCEAN model of personality. Like GPT-3, OCEAN is designed around a massive library of the English language. That should be reason enough to explore combining the two. How so? It’s simple. OCEAN further breaks down into two components per trait as follows, openness and intellect, industriousness and orderliness, enthusiasm and assertiveness, politeness and compassion, and finally, withdrawal and volatility. Straightforwardly, assign a value of 1–100 for each of the ten traits and the output will align with that of any English speaking person. Personality is a sliding scale based on frequencies. Adjusting the frequencies produces different personalities.
Technology lacks personality because we don’t have sufficient mental models, even for ourselves. In The Rosales Project, I detailed how much work it took to reach this point of understanding, and I’m giving the technical community the chance to stand on my shoulders, as well as the shoulders of others, who have done great work in fields such as language and psychology, which are becoming rapidly more relevant. MBTI, enneagram, and even Big 5 mental models that don’t allow for further breakdowns in the components just aren’t good enough to develop a strong understanding of personality or how to implement it from a mathematical perspective. What I’m recommending in this paper might be detailed enough to work.
What GPT-3 lacks is emotional intelligence. The paragraphs often come across as piles of nothing. Not only is the scope of information provided limited to what was originally input for the system to copy, meaning that the output is limited in terms of capacity for surprise, but the writing also fails to achieve any of the emotional appeal that even a 3rd grader could produce in their writing by virtue of being human. I’m suggesting a mechanism for humanization of this technology by introducing the audience to a mathematical and statistical basis for the emotional range we all exhibit in our use of language. Words only account for 7% of all communication among human beings, with tone and body language being far more important to us by nature. We are already limited to working with a fraction of what holds our attention when it comes to AI text generation, so every mental model we can use to improve it helps.
Machine learning algorithms can become equipped with personalities that are modelled after real human emotional ranges. OCEAN accounts for personality in both the macro and the micro. A person is both their overall frequency of emotion, and the way that those frequencies affect their slightest utterances. People exceptionally high in volatility use strong language and take longer to calm down. People who are low in enthusiasm are less upbeat or happy than those who are high, and are less likely to be in the mood to crack jokes. People who are low in compassion and politeness are less interested in the problems of others and more likely to be blunt and to interrupt. I see no reason that these same mental models can’t be applied to our AI for the purpose of humanizing a technology that we all see as potentially valuable.
In the next part, I will discuss some possible applications of the combination of this technology with OCEAN, as well as some of the practical and moral implications of doing this kind of work and research. Thanks for reading.
also bit shifts are fast.
